Low-Code Development Platforms (LCDPs) empower domain experts, potentially without software development training, to address their software requirements themselves by raising the level of abstraction beyond code. Domain experts typically interact with an LCDP through a graphical or text-based environment, which is more intuitive to use the more specific it is to their domain and to them as users. While most of the LCDPs from major vendors support a variety of users and domains by providing more general languages that require additional customization efforts, the modeling constructs for a specific LCDP must be created manually to support new domains and users - a laborious process. At the same time, Large Language Models (LLMs) demonstrate impressive code generation capabilities. However, domain experts often lack the skills to integrate the generated code into their organization’s software ecosystem.

Combining the power of both LCDPs and LLMs, this thesis focuses on automating the building of modeling constructs of LCDPs using LLMs (“LCDP GenAIrator”). Specific LCDPs allow domain experts to approach their tasks naturally because the modeling constructs mirror their understanding of the real world, while leveraging the code generation capabilities of LLMs provides a more efficient, automated way of creating LCDPs. More precisely, this thesis investigates how multimodal LLM input, including text (natural language), documents, structured data, images (e.g., screenshots), videos, code, and language specifications, can be used to create new LCDPs and extend existing software systems with LCDPs. Multimodal input is essential because a variety of artifacts contain domain knowledge, particularly modeling constructs. While their extraction can be automated, it requires different inputs that can be provided by domain experts. Additionally, the thesis explores how to utilize and optimize LLMs to excel at the task of generating modeling constructs from multimodal input using techniques such as fine-tuning, agent orchestration, and prompting strategies.

This thesis is carried out through industry-academia collaborations that allow to evaluate the “LCDP GenAIrator” using different scientific methods, from the lab to the industry partners’ sites.